Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

A Dual-Module Denoising Approach with Curriculum Learning for Enhancing Multimodal Aspect-Based Sentiment Analysis

T0 review · 5 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that a dual-module model, DualDe, is the first to concurrently denoise sentence-image and aspect-image noise in multimodal aspect-based sentiment analysis, and reports state-of-the-art F1 scores on Twitter-15 and…

desk verdict A competent engineering combo whose novelty and SOTA claims both hinge on one missing baseline, M2DF. read the letter →

arxiv 2412.08489 v1 pith:P6UJR3GI submitted 2024-12-11 cs.CV cs.MM

Nguyen Van Doan , Dat Tran Nguyen , Cam-Van Thi Nguyen This is my paper · ORCID
classification cs.CVcs.MM
keywords multimodalaspect-basedsentimentanalysissentence-imagedenoisingaspect-imagecurriculumlearningaspect-guidedattentionjointaspect-sentimentTwitter-15Twitter-17
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes DualDe, a model for multimodal aspect-based sentiment analysis that separates image noise into two kinds: noise at the level of the whole sentence-image pair, and noise at the level of individual image regions as they relate to a specific aspect. It claims to be the first to handle both kinds in a single system. A curriculum-learning module (HCD) sorts training samples by a composite difficulty score that mixes a fixed text-image similarity from a pretrained image-text model (CLIP) with a model-loss score refreshed each epoch, then feeds samples to the model in increasing difficulty. An aspect-enhance denoising module (AED) uses aspect-guided attention to suppress image regions and text tokens irrelevant to the aspect, and injects affective knowledge from SenticNet through a graph convolutional network. On the Twitter-15 and Twitter-17 benchmarks, the paper reports the highest precision, recall, and F1 among the compared models for the joint task, with JMASA F1 of 67.1 and 67.3.

What carries the argument

The central mechanism is the composite difficulty score $d^c_i = \alpha d^l_i + (1-\alpha) d^s_i$ with $\alpha = 0.8$, fused with a competence-based curriculum schedule $p(t)$ that admits samples only when $d^c_i < p(t)$. This carries the sentence-image denoising half. The aspect-image half is carried by the AESA and GCN sub-modules: an aspect-aware attention mechanism produces aspect-relevant hidden states, SenticNet values sharpen sentiment polarity, and a graph convolutional network over a weighted association matrix propagates affective dependencies across text and image nodes while suppressing unrelated visual blocks. Together these two sub-systems implement the paper's two-source denoising claim.

What would settle it

Run DualDe and the closest prior curriculum denoising framework on the same Twitter-15 and Twitter-17 splits and compare JMASA F1; if the prior framework already filters aspect-specific visual regions, or if it matches or beats the reported F1 of 67.1 and 67.3, the paper's novelty and state-of-the-art claims are refuted. As a second check, replace the fixed similarity score $d^s_i$ with random scores while keeping $d^l_i$; if the F1 does not drop, the claimed benefit of the hybrid difficulty metric is not causal.

Watch

Extended reading notes

Core claim

The paper's central claim is that sentence-image noise and aspect-image noise are distinct failure modes in multimodal aspect-based sentiment analysis, and that a model can address both simultaneously. DualDe combines the Hybrid Curriculum Denoising Module (HCD) and the Aspect-Enhance Denoising Module (AED). HCD orders training examples by a composite difficulty score $d^c_i = \alpha \cdot d^l_i + (1-\alpha) \cdot d^s_i$ with $\alpha = 0.8$, where $d^s_i$ is a fixed cosine similarity between CLIP text and image features and $d^l_i$ is a normalized per-sample generation loss refreshed each epoch; a competence function $p(t)$ admits only samples easier than the current competence. AED takes an aspect-aware attention output, adds SenticNet affective values, and runs a graph convolutional network over a weighted association matrix that encodes aspect-to-image and dependency-tree relations, so noisy visual blocks are filtered before the BART decoder predicts aspect spans and polarities. On Twitter-15 and Twitter-17, the paper reports JMASA F1 of 67.1 and 67.3 and states these are the highest scores across precision, recall, and F1 among the compared models.

Load-bearing premise

The paper's claims of novelty and state-of-the-art performance rest on the assumption that the closest prior curriculum denoising framework does not already handle aspect-image noise and therefore could be omitted from all comparisons; if that assumption is wrong, the claim to be the first to address both noise types and the claim of highest scores are unsupported.

Editorial extensions

If this is right

  • If the two-module design is correct, adding a sentence-level curriculum on top of aspect-level attention should beat either denoising strategy alone; the paper's ablation table supports this, with removal of HCD and of AESA both reducing F1.
  • The reported gains on two different tweet corpora imply the approach transfers across datasets with different noise profiles, not just a single benchmark.
  • The ratio test, which found $\alpha=0.8$ optimal, implies the model's own training signal is more informative than the fixed similarity score for ordering samples by difficulty.
  • The subtask results indicate the joint-task gains carry over to classification of known aspects (MASC) and extraction of aspect terms (MATE), although the MATE F1 gains are small.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension outside the paper: the same composite difficulty score could be applied to other multimodal sequence-to-sequence tasks, such as multimodal named entity recognition, where sentence-image noise is implicitly present.
  • Because the $\alpha=0.8$ weighting was selected by a search over ratios, a natural follow-up would be to anneal $\alpha$ during training, since the model-loss difficulty $d^l_i$ should become more reliable as training progresses.
  • The paper's distinction between the two noise types implies that region-level noise annotations would let the two modules be evaluated separately rather than only through ablations; without such labels, the relative contribution of each module is inferred rather than directly measured.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes DualDe, a framework for Multimodal Aspect-Based Sentiment Analysis (MABSA) that targets two types of visual noise: sentence-image noise and aspect-image noise. It combines a Hybrid Curriculum Denoising Module (HCD), which uses a weighted mixture of a CLIP-based similarity difficulty and a model-loss difficulty to schedule training, with an Aspect-Enhance Denoising Module (AED) built on a BART encoder-decoder with aspect-guided attention, SenticNet features, and a GCN over a dependency-based association matrix. Experiments on Twitter-15 and Twitter-17 report improved F1 over the listed baselines for JMASA, MASC, and MATE, and an ablation indicates that both main modules contribute. The central claims are that DualDe is the first model to concurrently address both noise types and that it achieves state-of-the-art precision, recall, and F1 on both benchmark datasets.

Significance. If the results were fully substantiated, the paper would make a useful incremental contribution by combining two types of denoising and by adapting curriculum learning with a composite difficulty metric. The ablation study and the positive F1 deltas on three tasks are encouraging, and the use of CLIP as an external similarity signal for curriculum difficulty is a reasonable design choice. However, the significance is currently capped by two load-bearing deficiencies: the most relevant prior framework, M2DF, is cited but entirely absent from the experiments, and the main curriculum weight α is selected on the test set. Because the paper's headline novelty and SOTA claims rest on these points, the significance of the reported numbers cannot be assessed as stated.

major comments (5)
  1. [§2.2, Tables 2–4] M2DF (Zhao et al., 2023), which the paper itself identifies in Section 2.2 as the closest curriculum-learning denoising framework for MABSA, never appears in any comparison table, and the only justification for its exclusion is the unsupported sentence 'they do not account for aspect-image noise.' The reference entry describes M2DF as a 'multi-grained multi-curriculum denoising framework,' which prima facie suggests it may already address both sentence-level and aspect-level noise. Without a direct comparison or at least a detailed, evidence-based discussion of M2DF's scope, the two central claims in Section 1 ('first to concurrently address both' noise types) and Section 4.3 ('highest scores across Precision, Recall, and F1') are unsubstantiated.
  2. [§3.1.3, §4.4.2, Fig. 4] The comprehensive difficulty weight α is selected by tuning on the test set: Section 3.1.3 states 'Empirical results indicate that setting α = 0.8 yields optimal performance,' and Section 4.4.2 evaluates candidate ratios directly by their test-set F1 scores. Because the reported JMASA F1 of 67.1 is the very number used to pick α, the result is an in-sample statistic rather than an honest evaluation. The paper does not describe a validation-based selection procedure or a nested resampling approach, so the reported improvements are optimistically biased and the SOTA claim is unsupported.
  3. [Tables 2–4] No standard deviations, confidence intervals, or significance tests are reported for any result, and many of the claimed improvements are smaller than 0.25 F1 points (e.g., 0.24 on Twitter-17 JMASA F1, 0.15 on Twitter-17 MATE F1). With single-run evaluations, such small deltas are within typical run-to-run variation for transformer-based models. The paper should report multiple runs with standard deviations and, ideally, statistical significance tests before claiming to outperform the previous state of the art.
  4. [§3.2.3, Eq. (16)] The GCN update rule is mathematically inconsistent as written. The summation runs over j = 1 to n, where n is the number of text tokens, but the association matrix A is defined in Section 3.2.2 as an (m+n) × (m+n) matrix over both image and text nodes. Additionally, the right-hand side uses h_i,l-1 inside the sum rather than h_j,l-1, which is the standard form. These issues make the graph convolution and its region-specific mixing ambiguous, and the implementation cannot be reproduced from the equations alone.
  5. [§3.1.4, §3.2.4, §4.1] Several critical hyperparameters are either not reported or not specified: the initial competence λ_init and the duration T in p(t) of Eq. (6) are never given numeric values; the fusion weights α1 and α2 in Eq. (17) are not stated to be learned parameters or fixed constants; and the dependency distance threshold in Section 3.2.2 is set to 2 without justification. Without these values, the HCD curriculum schedule and the AED fusion cannot be replicated, and the sensitivity of the results to these choices is unknown.
minor comments (5)
  1. [§1] The introduction contains a nearly verbatim duplicated sentence: 'In real-world scenarios, not all images are relevant to the accompanying text; some even mislead the contextual and emotional understanding of the sentence.' followed two sentences later by 'In real-world scenarios, images accompanying text may not always be relevant and can sometimes mislead the interpretation of the sentence’s context and emotion.' One of these should be removed.
  2. [§3.2.2, Eq. (14)] The rewritten matrix condition in Eq. (14) says A_ij = cos(h_i,h_j) under '(II) and (III)', but conditions (II) and (III) describe disjoint regions (text-to-image vs. text-to-text). The intended construction of the association matrix is unclear, in particular whether the aspect-driven entries apply to image nodes only or to both text and image neighbors.
  3. [§3.2.4, Eq. (19)] The notation H^T = W + ~H^T_2 introduces a subscript '2' that is undefined. It is also not clear whether W is the embedding matrix or a learned projection. Please clarify the dimensions and the intended role of each term.
  4. [§4.2] The list of compared models omits M2DF despite it being described as the closest related work; also, for the JMASA task, the table lists 'CLIP' as a baseline, but CLIP is primarily a feature extractor rather than a full MABSA model, and the manner in which it is applied to the joint task is not described.
  5. [Throughout] There are numerous typographical and stylistic errors, such as 'Diffculty' in the Section 3.1.2 heading, 'Aspect-Enhance' vs. 'Aspect-Enhanced' in module names, 'which unrelated to the specific aspects' in the abstract, and inconsistent capitalization of Twitter15/Twitter17. A careful proofreading pass is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the model's components are external or standard building blocks and the reported scores are empirical benchmark results, not quantities derived from their own definitions.

full rationale

The paper does not derive any prediction from its own inputs by construction. The HCD module's similarity difficulty uses cosine similarity between CLIP text and image features (Eqs. 1-2), an external pretrained model; the model-loss difficulty is the standard sequence loss (Eqs. 3-4), and the combined metric (Eq. 5) is a weighting of these two externally defined signals. Curriculum selection via p(t) (Eq. 6) follows Platanios et al. (2019). The AED module explicitly borrows the A3M aspect-attention mechanism from Zhou et al. (2023) and augments it with SenticNet affective values (Eqs. 7-13); the GCN propagation and BART decoder loss follow standard formulations. The main results in Tables 2-4 are empirical F1/accuracy measurements on Twitter-15 and Twitter-17, not quantities that reduce to the model's definitions or to a fitted parameter renamed as a prediction. The choice alpha=0.8 is a hyperparameter selected from the reported ablation (Fig. 4), which can raise a test-set-tuning concern, but it is not a logical circle: the reported F1 is a measurement of a trained model, not an algebraic consequence of the definition of alpha. The novelty and SOTA claims depend on the assertion that M2DF (Zhao et al., 2023) handles only sentence-image noise and on the absence of M2DF from the comparison tables; this is an evidentiary and baseline-completeness issue, not circularity. The only self-citation (Nguyen et al., 2024) appears in a general list of curriculum-learning applications and is not load-bearing. No load-bearing step reduces to its own input, so the paper is not significantly circular.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central results depend on a small number of hand-set hyperparameters (alpha=0.8, tree threshold=2), unreported curriculum parameters, and several unvalidated domain assumptions about CLIP similarity, loss-based difficulty, SenticNet, and the reused A3M module.

free parameters (4)
  • alpha (HCD difficulty mixing weight) = 0.8
    Chosen by tuning in Figure 4; balances model-loss difficulty dl and similarity difficulty ds in Eq. (5).
  • Curriculum schedule lambda_init and T = not reported
    Eq. (6) defines p(t) using lambda_init and T, but the paper never gives their values, so the curriculum schedule is not reproducible.
  • alpha1, alpha2 (AESA/GCN mixing weights) = not reported
    Eq. (17) blends AESA and GCN outputs; the scalar weights are learned or hand-set but their values and initialization are not given.
  • Dependency distance threshold = 2
    Text-text edges in the weighted association matrix are kept when dependency distance D_ij <= 2; chosen by hand, no sensitivity analysis reported.
assumptions (6)
  • domain assumption CLIP cosine similarity between sentence text and image is a valid measure of sentence-image noise/difficulty.
    Section 3.1.1, Eqs. (1)-(2); no validation that CLIP scores align with human noise labels.
  • domain assumption Sample loss under the current model is a reliable dynamic difficulty signal and is comparable across epochs after normalization.
    Section 3.1.2, Eqs. (3)-(4); loss can reflect model state but normalization by max is unstable if the max changes.
  • domain assumption Competence-based curriculum from Platanios et al. (2019) improves MABSA training.
    Section 3.1.4; adopted without independent validation for this task.
  • domain assumption Aspect-Aware Attention module from Zhou et al. (2023) effectively filters aspect-irrelevant visual blocks.
    Section 3.2.1; used as a component, not re-derived or benchmarked in isolation with error bars.
  • domain assumption SenticNet affective values carry useful sentiment priors for this task.
    Section 3.2.1, Eqs. (12)-(13); external knowledge assumed beneficial.
  • domain assumption GCN over a dependency-distance thresholded graph captures affective syntactic dependencies.
    Section 3.2.2-3.2.3; graph construction and GCN update assumed apt without ablation on graph type.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Dual-Module Denoising Approach with Curriculum Learning for Enhancing Multimodal Aspect-Based Sentiment Analysis." pith.science (2026). https://pith.science/paper/P6UJR3GI

@misc{pith2026241208489,
  author       = {Pith},
  title        = {Pith review of: A Dual-Module Denoising Approach with Curriculum Learning for Enhancing Multimodal Aspect-Based Sentiment Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P6UJR3GI}},
  note         = {Machine review of arXiv:2412.08489}
}
read the original abstract

Multimodal Aspect-Based Sentiment Analysis (MABSA) combines text and images to perform sentiment analysis but often struggles with irrelevant or misleading visual information. Existing methodologies typically address either sentence-image denoising or aspect-image denoising but fail to comprehensively tackle both types of noise. To address these limitations, we propose DualDe, a novel approach comprising two distinct components: the Hybrid Curriculum Denoising Module (HCD) and the Aspect-Enhance Denoising Module (AED). The HCD module enhances sentence-image denoising by incorporating a flexible curriculum learning strategy that prioritizes training on clean data. Concurrently, the AED module mitigates aspect-image noise through an aspect-guided attention mechanism that filters out noisy visual regions which unrelated to the specific aspects of interest. Our approach demonstrates effectiveness in addressing both sentence-image and aspect-image noise, as evidenced by experimental evaluations on benchmark datasets.

Figures

Figures reproduced from arXiv: 2412.08489 by the authors.

Figure 1
Figure 1. Illustration of Sentence-Image Denoising and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Model Overview After that, we normalized this difficulty score of i-th sample to [0.0, 1.0] by following formula: d l i = Li max 1≤j≤N Lj (4) where N is length of train dataset. Since the difficulty a batch sample (based on the loss metric) is entirely dependent on the model’s state, we update the loss metric at each epoch to ensure accurate evaluation. 3.1.3 Comprehensive Difficulty Metric The difficulty metric d s… view at source ↗
Figure 3
Figure 3. Illustrate the curve of the competence function [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Illustration Contribution Ratio Coefficient [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: The figure illustrates instances where sentence-image noise and aspect-image noise impact the effectiveness [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. CLAMP: Contrastive Learning with Adaptive Multi-loss and Progressive Fusion for Multimodal Aspect-Based Sentiment Analysis

    cs.CV 2025-07 conditional novelty 4.0 of 10

    CLAMP combines progressive attention fusion, multi-task contrastive learning, and uncertainty-based multi-loss weighting to report small F1 improvements over prior multimodal aspect-based sentiment analysis methods.

Reference graph

Works this paper leans on

32 extracted references · 20 canonical work pages · cited by 1 Pith paper

  1. [1]

    Yoshua Bengio, J \'e r \^o me Louradour, Ronan Collobert, and Jason Weston. 2009. https://api.semanticscholar.org/CorpusID:873046 Curriculum learning . In International Conference on Machine Learning

    2009

  2. [2]

    Yitao Cai, Huiyu Cai, and Xiaojun Wan. 2019. https://api.semanticscholar.org/CorpusID:196187316 Multi-modal sarcasm detection in twitter with hierarchical fusion model . In Annual Meeting of the Association for Computational Linguistics

    work page 2019

  3. [3]

    Cambria, Soujanya Poria, Rajiv Bajpai, and Bj \"o rn Schuller

    E. Cambria, Soujanya Poria, Rajiv Bajpai, and Bj \"o rn Schuller. 2016. https://api.semanticscholar.org/CorpusID:2684375 Senticnet 4: A semantic resource for sentiment analysis based on conceptual primitives . In International Conference on Computational Linguistics

    work page 2016

  4. [4]

    Guimin Chen, Yuanhe Tian, and Yan Song. 2020. https://api.semanticscholar.org/CorpusID:227231566 Joint aspect extraction and sentiment analysis with directional graph convolutional networks . In International Conference on Computational Linguistics

    work page 2020

  5. [5]

    Tao Chen, Damian Borth, Trevor Darrell, and Shih-Fu Chang. 2014. https://api.semanticscholar.org/CorpusID:16645406 Deepsentibank: Visual sentiment concept classification with deep convolutional neural networks . ArXiv, abs/1410.8586

    arXiv 2014

  6. [6]

    Minghao Hu, Yuxing Peng, Zhen Huang, Dongsheng Li, and Yiwei Lv. 2019. https://api.semanticscholar.org/CorpusID:182952458 Open-domain targeted sentiment analysis via span-based extraction and classification . ArXiv, abs/1906.03820

    arXiv 2019

  7. [7]

    Xincheng Ju, Dong Zhang, Rong Xiao, Junhui Li, Shoushan Li, Min Zhang, and Guodong Zhou. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.360 Joint multi-modal aspect-sentiment analysis with auxiliary cross-modal relation detection . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4395--4405, Online and Pu...

    doi:10.18653/v1/2021.emnlp-main.360 2021

  8. [8]

    Zaid Khan and Yun Raymond Fu. 2021. https://api.semanticscholar.org/CorpusID:236912847 Exploiting bert for multimodal target sentiment classification through input space translation . Proceedings of the 29th ACM International Conference on Multimedia

    work page 2021

Show all 32 references
  1. [9]

    Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdel rahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2019. https://api.semanticscholar.org/CorpusID:204960716 Bart: Denoising sequence-to-sequence pre-training for natural language generation, trans...

    2019

  2. [10]

    Yan Ling, Jianfei Yu, and Rui Xia. 2022. https://api.semanticscholar.org/CorpusID:248227577 Vision-language pre-training for multimodal aspect-based sentiment analysis . In Annual Meeting of the Association for Computational Linguistics

    2022

  3. [11]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. https://api.semanticscholar.org/CorpusID:198953378 Roberta: A robustly optimized bert pretraining approach . ArXiv, abs/1907.11692

    2019 arXiv

  4. [12]

    Jinliang Lu and Jiajun Zhang. 2021. https://api.semanticscholar.org/CorpusID:237605036 Exploiting curriculum learning in unsupervised neural machine translation . In Conference on Empirical Methods in Natural Language Processing

    2021

  5. [13]

    Cam-Van Thi Nguyen, Cao-Bach Nguyen, Duc-Trong Le, and Quang-Thuy Ha. 2024. https://aclanthology.org/2024.lrec-main.380 Curriculum learning meets directed acyclic graph for multimodal emotion recognition . In Proceedings of the 2024 Joint International Conference on Computatio...

    2024

  6. [14]

    Emmanouil Antonios Platanios, Otilia Stretcu, Graham Neubig, Barnab \'a s P \'o czos, and Tom Michael Mitchell. 2019. https://api.semanticscholar.org/CorpusID:85498775 Competence-based curriculum learning for neural machine translation . ArXiv, abs/1903.09848

    2019 arXiv

  7. [15]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. https://api.semanticscholar.org/CorpusID:231591445 Learning transferable visual model...

    2021

  8. [16]

    Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. https://api.semanticscholar.org/CorpusID:160025533 Language models are unsupervised multitask learners

    2019

  9. [17]

    Lin Sun, Jiquan Wang, Kai Zhang, Yindu Su, and Fangsheng Weng. 2021. https://api.semanticscholar.org/CorpusID:231839433 Rpbert: A text-image relation propagation-based bert model for multimodal ner . ArXiv, abs/2102.02967

    2021 arXiv

  10. [18]

    co-curricular learning

    Wei Wang, Isaac Caswell, and Ciprian Chelba. 2019. https://api.semanticscholar.org/CorpusID:174798125 Dynamically composing domain-data selection with clean-data selection by “co-curricular learning” for neural machine translation . ArXiv, abs/1906.01130

    2019 arXiv

  11. [19]

    Hanqian Wu, Siliang Cheng, Jingjing Wang, Shoushan Li, and Lian Chi. 2020 a . https://api.semanticscholar.org/CorpusID:222210395 Multimodal aspect extraction with region-aware alignment network . In Natural Language Processing and Chinese Computing

    2020

  12. [20]

    Cai, Junying Chen, Ho fung Leung, and Qing Li

    Zhiwei Wu, Changmeng Zheng, Y. Cai, Junying Chen, Ho fung Leung, and Qing Li. 2020 b . https://api.semanticscholar.org/CorpusID:222278618 Multimodal representation with embedded visual guiding objects for named entity recognition in social media posts . Proceedings of the 28th...

    2020

  13. [21]

    Bo Xu, Shizhou Huang, Ming Du, Hongya Wang, Hui Song, Chaofeng Sha, and Yanghua Xiao. 2022. https://api.semanticscholar.org/CorpusID:252819503 Different data, different modalities! reinforced data splitting for effective multimodal information extraction from social media post...

    2022

  14. [22]

    Hang Yan, Junqi Dai, Tuo Ji, Xipeng Qiu, and Zheng Zhang. 2021. https://api.semanticscholar.org/CorpusID:235367681 A unified generative framework for aspect-based sentiment analysis . ArXiv, abs/2106.04300

    2021 arXiv

  15. [23]

    Jianfei Yu and Jing Jiang. 2019. https://api.semanticscholar.org/CorpusID:199465957 Adapting bert for target-oriented multimodal sentiment classification . In International Joint Conference on Artificial Intelligence

    2019

  16. [24]

    Jianfei Yu, Jing Jiang, and Rui Xia. 2020 a . https://api.semanticscholar.org/CorpusID:209459462 Entity-sensitive attention and fusion network for entity-level multimodal sentiment classification . IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:429--439

    2020

  17. [25]

    Jianfei Yu, Jing Jiang, Li Yang, and Rui Xia. 2020 b . https://api.semanticscholar.org/CorpusID:220045906 Improving multimodal named entity recognition via entity span detection with unified multimodal transformer . In Annual Meeting of the Association for Computational Linguistics

    2020

  18. [26]

    Jianfei Yu, Jieming Wang, Rui Xia, and Junjie Li. 2022. https://doi.org/10.24963/ijcai.2022/622 Targeted multimodal sentiment classification based on coarse-to-fine grained image-target matching . In Proceedings of the Thirty-First International Joint Conference on Artificial ...

    2022 doi

  19. [27]

    Dong Zhang, Suzhong Wei, Shoushan Li, Hanqian Wu, Qiaoming Zhu, and Guodong Zhou. 2021. https://api.semanticscholar.org/CorpusID:235363807 Multi-modal graph fusion for named entity recognition with targeted visual guidance . In AAAI Conference on Artificial Intelligence

    2021

  20. [28]

    Lei Zhang and B. Liu. 2012. https://api.semanticscholar.org/CorpusID:38022159 Sentiment analysis and opinion mining . In Synthesis Lectures on Human Language Technologies

    2012

  21. [29]

    Fei Zhao, Chunhui Li, Zhen Wu, Yawen Ouyang, Jianbing Zhang, and Xinyu Dai. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.561 M 2 DF : Multi-grained multi-curriculum denoising framework for multimodal aspect-based sentiment analysis . In Proceedings of the 2023 Conference ...

    2023 doi

  22. [30]

    Ru Zhou, Wenya Guo, Xumeng Liu, Shenglong Yu, Ying Zhang, and Xiaojie Yuan. 2023. https://api.semanticscholar.org/CorpusID:259064036 Aom: Detecting aspect-oriented information for multimodal aspect-based sentiment analysis . In Annual Meeting of the Association for Computation...

    2023

  23. [31]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  24. [32]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.